Rate-Aware KAN-Mamba Framework for Data-Driven Tool Wear State Prediction from Multi-Rate Spindle Sensor Stream

Abstract

This paper proposes a rate-aware KAN-Mamba framework for data-driven tool wear state prediction from multi-rate spindle sensor streams. The core methodological contributions center on three tightly integrated components that foreground sequential modeling, multi-rate sensor fusion, and efficient long-sequence engineering rather than machining-specific heuristics. First, rate-aware tokenization divides asynchronous spindle channels into independent segments according to their native sampling rates, preserving both high-frequency transient events and low-frequency supervisory evidence. Second, a KAN-based nonlinear feature mapping learns conditional local basis functions that align processing-state interactions with sensor readings, enabling adaptive non-linear transformation across varying regimes. Third, a Mamba-style selective state update module compresses long-term temporal memory while applying context-gated updates, capturing degradation sequences with minimal computational overhead. Extensive experiments on industrial spindle sensor streams demonstrate substantial practical gains: improved RMSE and MAE under transient and severe-wear regimes, transition-state tracking accuracy, and robustness to missing or noisy channels. Compared with LSTM, GRU, and Transformer baselines, the framework outperforms by more than 20 % in MAE while achieving an 18.6-minute early-warning lead time and 18.9 ms inference latency under robustness tests. By foregrounding rate-aware multi-rate sensor fusion and KAN-Mamba sequential modeling, the approach supplies an efficient, transferable solution for industrial time-series analytics and aligns directly with data-centric engineering objectives for smart manufacturing decision support.